_critic
HarnessX Critic (Tier 2.1) — adversarial review of evolved-skill proposals against trace evidence. Detects reward hacking and manifest/evidence contradictions. Out-of-band LLM counterpart to the in-loop deterministic critic in src/evolve/critic.rs.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 1.6k
- Forks
- 491
01 前言 Front Matter
- name
- _critic
- description
- HarnessX Critic (Tier 2.1) — adversarial review of evolved-skill proposals against trace evidence. Detects reward hacking and manifest/evidence contradictions. Out-of-band LLM counterpart to the in-loop deterministic critic in src/evolve/critic.rs.
03 SKILL.md 正文
Critic — Adversarial Proposal Review
In-loop vs out-of-band. epic-harness forbids external LLM calls from production, so the reflect loop ships a deterministic critic (
src/evolve/critic.rs) that gates seeding when reward hacking is suspected. THIS skill is the out-of-band LLM version a meta-agent or human runs during/evolvereview for the cases the deterministic check cannot catch (non-local effects, manifest/evidence nuance).
When to Trigger
- During
/evolvereview of newly seeded skills - When
reward_hacking_suspectedis true in metrics - After a seesaw-regression round, before re-proposing
Process
1. Gather the proposal + evidence
- Read the evolved skill proposal(s) from this round
- Read the EditManifest (edit_type, target, intended_effect, predicted_impact)
- Read the round's TaskDigests (outcome, failure_categories, evidence_excerpts)
- Read recent score_history dimension_averages (tool_success, output_quality, execution_cost)
2. Falsify the manifest (paper §4.3, Table 9)
For each proposal, ask: does the trace evidence support the predicted_impact?
- If the manifest claims "Lift avg_score_with" but output_quality is regressing → Reject
- If the manifest claims a tool fix but the implicated component shows no change → Warn
- If the evidence corroborates the predicted effect → Approve
3. Reward-hacking cross-check
- Is execution_cost rising while output_quality falls across the window?
- Could the skill be gaming a metric (fewer tool calls inflating cost score) rather than improving outcomes?
- If yes, the skill must NOT ship — flag for the rejected buffer.
4. Non-local effect scan
- Will this skill's guard rules conflict with existing skills (overlapping triggers, contradictory rules)?
- Does it interact with shared state (context, memory, control) in a way the manifest didn't account for?
Anti-Rationalization
| Excuse | Rebuttal | Do instead |
|---|---|---|
| "The score went up, so it works" | Score can rise via metric gaming | Verify the outcome improved, not just the score |
| "The seesaw passed, it's safe" | Seesaw is coarse; sub-threshold coupling evades it | Check dimension deltas, not just aggregate pass |
| "It's just a prompt tweak" | Prompt edits have non-local effects on shared context | Trace the effect across skills, not just the target |
Evidence Required
- [ ] Manifest's predicted_impact checked against observed dimension deltas
- [ ] reward_hacking_suspected consulted
- [ ] No conflict with ≥1 existing skill demonstrated
- [ ] Verdict (Approve/Warn/Reject) recorded per proposal with reason
Red Flags
- Approving a skill whose only evidence is "score went up"
- Ignoring a falling output_quality because execution_cost rose
- Shipping after a seesaw-regression round without explicit justification
- Treating the deterministic critic as sufficient for non-local effects (it is not — it only checks reward hacking + score-claim contradiction)
04 同仓库更多 Skills
worktable-review-with-annotations
Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.
worktable-find-and-synthesize-context
Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.
worktable-create-or-update-docs
Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.
worktable-create-or-manage-records
Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.
worktable-create-or-manage-interactive-html
Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.
worktable-collaborate-threads
Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.